Reasoning Config
A reasoning config is the setting used to determine how extensively an AI model is allowed to think before giving its answer. It thus directly controls how good, how fast, and how expensive the answer turns out to be.
Modern language programs like ChatGPT can answer a question immediately or first spend a while computing internally. This intermediate computation is a kind of invisible scratchpad: the program tries out solution paths, discards them, and in the end writes down only the result. How long this scratchpad may be is not fixed in advance. It can be specified when calling the model, and this exact specification is called a reasoning config. Roughly translated, it means “setting for thinking.” It doesn’t say WHAT the program should think about, only HOW MUCH effort it may put into it.
The price of thinking
Every computational step of an AI model costs electricity, processing time, and thus money. Anyone who turns up the thinking level gets noticeably better results on hard tasks. For mathematics, programming, or multi-step logic questions, the hit rate can increase perceptibly. In exchange, an answer no longer takes one second, but half a minute or longer.
Conversely, full-on thinking is pure waste for simple tasks. For the question “What’s the weather like in Hamburg?” nobody needs a long chain of thought. Companies that process millions of requests every day save substantial sums with the right setting. That’s why the reasoning config is a recurring topic in business news about AI providers: it is the most direct lever between quality and cost.
A common misconception: more thinking doesn’t make a model smarter than it actually is. It only extracts more from what the model has already learned. Knowledge that was missing during training doesn’t materialize even after two minutes of computing.
From “low” to “high”: what you actually configure
In practice, a reasoning config is often just a single word in the API call. Common levels include “minimal,” “low,” “medium,” and “high.” Some providers instead expect a number: a budget of so-called tokens, i.e., text chunks, that the scratchpad may contain at most. A budget of 1000 tokens corresponds to roughly one book page’s worth of internal notes.
Technically, this isn’t a switch that unlocks new capabilities. The model was prepared during training to stretch out its thinking to varying lengths. The config specifies how many of these steps it’s allowed to take before it must move on to the final answer. If the budget runs out, the model breaks off its chain of thought and answers with whatever it has arrived at by that point.
It’s important to distinguish this from the prompt. The prompt is the task you pose. The reasoning config is a setting alongside it, similar to the video quality of a video. Both together produce the result, but they operate at completely different points.
Where the thinking level shows up
In the chat window, you encounter this principle as a selection list of models. When a provider lets you choose between a fast variant and a “thinking” variant, the same technique is often behind it. Some interfaces additionally display a note like “Thinking…” along with a summary of the intermediate steps.
The reasoning config becomes truly visible for developers. Anyone who integrates an AI model via an API sets it individually per request. A customer service bot runs at a low level during the day, while an overnight contract analysis runs at a high level. This fine-grained control is one of the reasons why the same AI can seem brilliant in one product and disappointing in another.
In the news, the term is usually encountered in connection with price lists and benchmarks, i.e., comparative tests of models. The results of such tests are only comparable if all models were tested with the same thinking level. Providers therefore now also state which setting was used to achieve a record score.